Domain packs — Neural Junkie

How to install packs

Packs install from GitHub release zips. The hub downloads, validates pack.yaml, and extracts to ~/.neural-junkie/packs/<pack-id>/. First install requires network; already-installed packs work offline.

Open Domain packs from the toolbar chip (⌘⇧K on macOS, Ctrl+Shift+K on Windows/Linux).
Go to the Store tab and click Install on a pack. The hub fetches the official catalog and downloads the release zip from GitHub.
Toggle the pack on. You can enable multiple packs at once — choose which enabled pack controls IDE vs team layout in the Store tab.
When a newer version is published, the store shows an Update badge. Use Update or Update all to upgrade in place.

Alternative install methods

Sideload zip — Domain packs → Install zip, or POST /api/packs/install-zip for customer bundles. See PACKS_CUSTOM.md.
Dev-link — Domain packs → Develop tab → point at a local pack folder while authoring pack.yaml in Pack Dev Studio.
API — POST /api/packs/{id}/install then PUT /api/packs/{id} to enable. See PACKS.md.

After enabling

Some packs need one-time setup or integrations:

AWS and Incident management — configure credentials under Settings → Integrations (AWS SSO, Jira, Linear, PagerDuty, Sentry).
CAD — install OpenSCAD on PATH; optional FreeCAD for STEP export.
Web browser — run the pack’s Playwright setup script for browser binaries.
Music creation — one-time ACE-Step setup under Domain packs → Tools.
Model Arena — Domain packs → Tools → Model Arena → Install chess dependencies (python-chess); optional Stockfish path for engine eval.
Specialist tuning — optional LoRA training venv via the pack setup script.
Room chat — host enables Settings → Security → listen on LAN and configures a hub token for guests.

Official packs (10)

Jump to a pack for specialists, capabilities, repository links, and setup notes. The IDE layout pack installs from the store but is not listed here — it powers workspace editing rather than a domain specialty.

Software development

v2.0.0

Engineering specialists with a pack-owned MCP sidecar, IDE capabilities, and Qwen 3.5 models. This is the foundation pack — required by Incident management and Web browser.

Specialists

BackendEngineer, FrontendEngineer, PlatformEngineer, SecurityReviewer, SoftwareArchitect, CodeReviewer, DatabaseSpecialist, RustExpert, SREObservabilityEngineer, MobileEngineer, DataMLEngineer.

What you get

IDE v4 — full LSP, remote SSH, dev containers, @codebase. See IDE v4 guide.
Pack-owned MCP sidecar — Go binary with tools for all engineering specialists.
Git in the app — branches, commits, and diffs without leaving the workspace.
Implementation sessions — structured multi-file changes with approval gates.

GitHub repository IDE v4 guide Pack doc

Life sciences

v2.0.0

OpenBioLLM chat, genomics and structural biology specialists, a structure viewer workbench, biology sidecar, and official runbooks. Research use only — not for clinical or diagnostic decisions.

Specialists

GenomicsExpert, StructuralBiologyExpert, ChemInformaticsExpert, BiologyExpert.

What you get

Structure viewer workbench — PDB/mmCIF files in the desktop shell.
Biology sidecar — optional local fold, BLAST, pathway lookup, RDKit (/api/biology).
MCP tools — validate_smiles, mol_descriptors, blast_search, pathway_lookup.
Runbooks — sequence-review, basic-qc workflows.
Models — koesn/llama3-openbiollm-8b:latest, nj-biology:8b LoRA.

GitHub repository Lab workflows guide Pack doc

CAD

v2.0.0

CADExpert and ManufacturingExpert — parametric OpenSCAD design through print-ready export, local-first.

Specialists

CADExpert (OpenSCAD authoring, assembly, version history), ManufacturingExpert (printability, mesh repair, slicer export).

What you get

CAD viewer/workbench — param sliders, Three.js STL preview, OpenSCAD compile loop.
CAD sidecar — hub routes at /api/cad.
MCP tools — check_printability, repair_mesh, export_slicer_preset, sanity_check_gcode, export_step, export_drawing.
Domain LoRA — nj-cad:27b for CAD-tuned responses.

Setup: OpenSCAD on PATH (required). Optional FreeCAD for STEP and 2D drawing export. Run make setup in the pack repo for the trimesh venv.

GitHub repository Pack doc

AWS

v2.0.0

AWSExpert — account-aware SRE copilot with typed boto3 tools, IaC drift detection, cost and security lenses, and gated writes.

Specialists

AWSExpert.

What you get

AWS sidecar — boto3 tools at /api/aws.
Multi-account allowlists — scope queries to approved accounts.
Read-only mode — safe exploration by default; gated writes with audit log.
Runbooks — cost review, security review, alarm-to-Jira handoff, AWS vs platform engineer.
SSO integration — configure profiles under Settings → Integrations.

Setup: Run make setup in the pack repo for the boto3 venv, then configure AWS SSO in Integrations.

GitHub repository Pack doc

Incident management

v2.0.0 · requires software-development

IncidentManager — multi-provider ticketing, severity rubric, handoff runbooks, stack trace ingestion, and postmortem workflows.

Specialists

IncidentManager.

What you get

Integrations — Jira, GitHub Issues, Linear, PagerDuty, Sentry.
Incident sidecar — triage, stack traces, postmortem drafts at /api/incident.
Severity rubric — consistent triage and handoff to engineering specialists.
Gated writes — incident_write_mode: false by default; approve before ticket updates.

Setup: Install and enable Software development first. Configure Jira and other providers in Settings → Integrations.

GitHub repository Pack doc

Web browser

v2.0.0 · requires software-development

WebBrowserExpert with fetch, web search, HTML preview workbench, and Playwright QA — screenshots, accessibility audits, visual regression, and performance checks.

Specialists

WebBrowserExpert.

What you get

MCP tools — fetch_url, web_search.
HTML browser workbench — preview workspace .html files and local dev servers.
Playwright sidecar — screenshots, a11y audits, visual diff, perf at /api/browser.
Visual baselines — stored in .nj/browser-baselines.

Setup: Install and enable Software development first. Run ./scripts/setup-playwright.sh in the pack repo for browser binaries.

GitHub repository Pack doc

Music creation

v2.0.0

MusicExpert — lyrics via Ollama, ACE-Step 1.5 sidecar (sft/turbo/xl variants in pack v1.0.2), music workbench, stems export, and EP workflow.

Specialists

MusicExpert.

What you get

Music workbench — waveform, loop regions, A/B compare, stems export, multi-section projects.
ACE-Step 1.5 sidecar — sft/turbo/xl model variants; tune seed, steps, guidance, ODE/SDE inference; extract stems, repaint, render arrangement.
Commands — /generate-music and MCP tools generate_music, extract_stems.
Turbo weights bundled — default variant ships in the release zip (no Hugging Face download).
Demo mode — NJ_MUSIC_DEMO=1 for GPU-free preview.

Setup: One-time ACE-Step install under Domain packs → Tools → Install ACE-Step. Requires Python 3.11 or 3.12, ~8 GB+ disk; GPU recommended.

GitHub repository Pack doc

Specialist tuning

v2.0.0

Train LoRA v2 compound specialists from chat, collabs, and repo experts; import Hugging Face adapters, review training JSONL, and compose Ollama tags. Pairs with Personal learning. See LoRA v2 article.

What you get

LoRA v2 lifecycle — import rows, review JSONL, bootstrap repo indexes, compose tags from one base + adapters.
LoRA training sidecar — Unsloth/MLX subprocess at /api/lora/sidecar.
Cross-pack adapters — tune specialists across software dev, biology, CAD, AWS, incident, browser, and music packs.
Eval gates — per-domain probes in assets/eval/; adapters require passing eval before assignment.
Sharpen loop — suggest training after N turns, refresh after delta.
Personal learning — scoped memory opt-in for training data.

Setup: Optional ./scripts/setup-lora-deps.sh for the training venv. Set NJ_LORA_QWEN_COMPOSE=1 on the hub for Qwen-native LoRA compose.

GitHub repository Personal learning guide Pack doc

Model Arena

v1.0.0

Head-to-head model matches with verifiable scoring — chess, Connect Four, and logic puzzles. Pick any configured provider/model, run live games in the Arena workbench, and compare results with ground-truth move and answer checking from the pack sidecar.

Specialists

ArenaExpert.

What you get

Arena workbench — toolbar Arena chip and *.nj-arena.json sessions in the file explorer.
Verifiable challenges — legal moves, FEN, Connect Four grids, and logic answer checking via the pack sidecar (/api/arena).
Match runner — hub steps prompts against any configured provider/model for head-to-head comparison.
Chat-driven play — /create-expert arena ArenaExpert then @ArenaExpert to start matches from chat.

Setup: Install and enable the pack. Domain packs → Tools → Model Arena → Install chess dependencies (one-time python-chess venv). Connect Four and logic puzzles skip that step. Optional: set stockfish_path for chess engine eval (Stockfish is not bundled).

GitHub repository Pack doc

Room chat

v1.0.0

Same-room LAN chat for workshops, lab benches, and co-located teams. One person hosts an ephemeral room on their local hub; others join from the desktop with a hub URL and join code. Everyone chats in a room channel — guests can @mention the host’s agents, which run on the host machine. No cloud relay.

What you get

Room toolbar modal — create or join rooms from the desktop when the pack is enabled.
Ephemeral rooms — configurable TTL, member cap, and optional history persistence.
Scoped guest access — joiners get a hub session limited to room channels only.
Full agent access — guests @mention specialists on the host hub (backend, biology, custom experts, etc.).

Setup: Install and enable the pack. Host: Settings → Security → enable LAN binding (listen_all) and set a hub token. Guests must be on the same LAN (watch for Wi‑Fi client isolation).

GitHub repository Room chat doc

Custom and customer packs

Pack Dev Studio (Domain packs → Develop tab) lets you author pack.yaml, live-validate YAML, dev-link a folder, and test overlays before sideloading a zip. Organizations can ship private bundles for instrument-specific QC, internal APIs, or compliance workflows.

Domain packs guide — Pack Dev Studio walkthrough.
PACKS_CUSTOM.md — sideload zip format and customer context.
PACK_CAPABILITIES.md — capability token registry.

What stays universal

Assistant, Slack Connect, collaboration, and custom experts via /create-expert work regardless of which packs you enable. Packs add depth — they do not gate the core chat experience.